You can deploy this model and start earning money today!

ridhwan99/
bert-sentiment-healthcare

Name: bert-sentiment-healthcare
Availability: InStock

safetensors

bert

mit

Views

No views yet

BERT Sentiment Classifier for Patient Health Risk

Fine-tuned bert-base-uncased for 3-class sentiment classification on patient questionnaire responses. The model predicts whether the patient's self-reported symptoms indicate positive, neutral, or negative health sentiment.

Labels

Label	Description
positive	No abnormal symptoms detected; patient reports "mild fatigue, stable breathing, no severe discomfort"
neutral	Some symptoms reported but tagged as low concern / low deterioration risk
negative	High concern symptoms affecting daily activities

Dataset

Source: patient_dataset.csv (2,000 synthetic patient records)
Input: questionnaire_text column — patient symptom descriptions with giveaway phrases (concern level, daily-activity impact) stripped to prevent label leakage
Labels: Derived from deterioration_risk (0/1) plus text parsing for the 0→positive/neutral split
Split: 80% train (1,600), 20% test (400), stratified

Training Details

Config	Value
Base Model	`bert-base-uncased`
Epochs	30
Batch Size	16
Learning Rate	2e-5
Optimizer	AdamW
Max Sequence Length	128 tokens
Class Weights	Inverse frequency (negative class weighted ~2.9x)
Threshold Tuning	Negative-class logits boosted by ~3.0x at inference

Performance (Test Set)

Class	Precision	Recall	F1-Score	Support
positive	1.00	1.00	1.00	177
neutral	0.92	0.99	0.95	177
negative	0.94	0.67	0.78	46
accuracy			0.96	400
macro avg	0.95	0.89	0.91	400

Training Loss

Usage

python

from transformers import BertTokenizer, BertForSequenceClassification
import torch

tokenizer = BertTokenizer.from_pretrained("models/bert_patient_risk_sentiment")
model = BertForSequenceClassification.from_pretrained("models/bert_patient_risk_sentiment")

text = "The patient reports shortness of breath, fever and body aches."
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True, max_length=128)

with torch.no_grad():
    logits = model(**inputs).logits
    # Apply tuned threshold for negative class
    logits[:, 2] *= 3.0
    pred = torch.argmax(logits, dim=1).item()

label_map = {0: "positive", 1: "neutral", 2: "negative"}
print(label_map[pred])

ridhwan99/bert-sentiment-healthcare